Can We Spot the Fakes? Detecting Political News Deepfakes from Different AI Models

Introduction This thesis examines the potential of Generative Pre-trained AI (GPAI) models to create deepfake news, particularly in political contexts. As AI technology progresses, understanding its capacity for content manipulation becomes crucial for developing effective regulatory frameworks like the EU AI Act. Research Question There is a need for a deeper understanding of the potential of GPAI models to manipulate content. Therefore, this study explores how different AI models vary in their ability to create deepfake news and how effective these models are at fostering disinformation. This thesis aims to answer the following question: How difficult is it to detect political news deepfakes generated by different AI models? Method This research uses a mixed-methods approach. First, an experiment is conducted to assess the capabilities of various AI models in generating deepfake political news articles. This is followed by a survey with the purpose of assessing how believable these articles are, and determining how likely they are to be used to spread false or misleading information (disinformation). The data analysis combines descriptive and inferential statistical methods to draw comprehensive conclusions. Results The survey showed that most respondents (60 percent) used AI tools very often and rated themselves as confident in handling AI-generated content. Participants achieved higher detection accuracy with images than videos, a statistically significant difference confirmed by a McNemar test. While the majority (30) found the media content credible, a notable number (12) expressed skepticism. Overall, results highlight greater difficulty in identifying AI-generated videos even for a confident, tech-savvy, and predominantly young adult population. Conclusion Regardless of the level of experience with AI tools, end users can struggle to identify synthetic media, especially videos, highlighting the deceptive realism of current generative models. The results emphasize the critical role of the media format in shaping perceived credibility and detectability. As video deepfakes grow more sophisticated, there is an urgent need for public education and policy makers to mitigate misinformation and protect democratic processes.